Benchmarking HEp-2 cell segmentation methods in indirect immunofluorescence images - standard models to deep learning

Abstract

Indirect Immunofluorescence (IIF) stained Human Epithelial (HEp-2) cells are considered the gold standard for detecting autoimmune diseases. Accurate cell segmentation, though often viewed as an intermediary step to downstream tasks like classification, significantly enhances overall performance when executed with precision. In this study, we conduct a systematic literature review of HEp-2 cell segmentation techniques, identifying 28 key papers utilizing traditional image processing, machine learning classifiers, deep convolutional neural networks (CNNs), and generative adversarial network (GAN) frameworks. Building on these insights, we benchmark 17 CNN models without pretraining and 8 CNN models pretrained on ImageNet using both Frozen Encoder and Tunable Encoder strategies on the I3A dataset. Cross-validation (CV) and Benjamini–Hochberg (BH) significance correction were employed to ensure statistical rigor in model comparisons. Domain-Specific Pretraining (DSPT) experiments demonstrated performance improvements, particularly for underrepresented classes, while Data Augmentation strategies (DA-1 and DA-2) revealed distinct impacts across model categories. GAN-based segmentation experiments using the top-performing CNN architectures as generators within a Pix2Pix framework revealed performance degradation due to data limitations and adversarial training instabilities. Nonetheless, GANs displayed class-specific improvements in visual alignment of segmentation masks. Results were evaluated comprehensively across eight performance metrics, including Dice, IOU, Accuracy, Precision, Sensitivity, Specificity, AU-ROC and AU-PR. This work offers a robust benchmarking of state-of-the-art CNN, GAN, and Transformer-based models for HEp-2 cell segmentation, providing valuable insights for future research directions, including ensemble approaches, dynamic patch sampling, and diffusion models.